What’s New in Dementia Risk Prediction Modelling? An Updated Systematic Review Open Access

As strategies to tackle the global burden of disease associated with dementia move towards a greater focus on primary prevention, in particular risk reduction, the identification of individuals at high risk has become extremely important. Such an approach typically combines known risk factors into an algorithm to calculate the probability of an individual developing incident disease. Such strategies work well in areas such as cardiovascular disease [1, 2] and stroke [3, 4] prevention. Despite dementia being a leading cause of disability and death on a global scale [5], there is no recommended tool to identify individuals at increased risk who might benefit from early intervention to delay or prevent it.

While screening for dementia risk is not currently recommended [6, 7], numerous prediction models have been developed. Indeed, previous systematic reviews [8‒10] have identified over 100 different models, largely developed in high-income White populations with variable discriminative performance (c-statistic range 0.49–0.89) and over a range of follow-up (1 year to >20 years). Where models have been externally validated the results are mixed [8, 9] and insufficiently reported [11]. New models are constantly being developed, including some that take advantage of more advanced computational techniques such as artificial intelligence (AI), in particular, machine learning (ML) algorithms.

Therefore, in this systematic review, we provide an update, covering all literature published after our 2015 review [9] on current developments in dementia risk prediction modelling. We aim to (i) synthesize the literature on all new dementia risk prediction models including reporting their discriminative accuracy and calibration where available, and external validation results; (ii) identify key risk factors for dementia incorporated into the different models; and (iii) by combining the results with our two previous reviews [8, 9] to make policy recommendations on dementia risk prediction. These recommendations are necessary to facilitate and coordinate ongoing and future dementia preventative work.

The protocol was registered on PROSPERO (Reference CRD42022320630). The systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) guidelines (online suppl. material 1; for all online suppl. material, see https://doi.org/10.1159/000539744) [12].

Embase (via Ovid), Medline (via Ovid), Scopus, and Web of Science were searched from the final date of our previous review (March 14, 2014) to the June 10, 2022. The following concepts were included in the search strategy and mapped to Medical Subject Headings (MeSH) where appropriate: dementia, Alzheimer disease, predict, develop, incident, sensitivity, specificity, ROC, area under the curve (AUC), and concordance statistic (online suppl. material 2). The electronic search results were transferred to Endnote software and de-duplicated [13]. The reference lists of all included articles were manually checked to ensure relevant studies were not missed in the electronic search.

Studies were selected for inclusion based on three criteria. Firstly, the study sample had to be population- or community-based (including electronic health record datasets) and examine risk of incident dementia in later life (i.e., at ≥60 years of age). Cross-sectional, case-control, trials, and clinical-based studies were excluded. There was no restriction on baseline age or follow-up time, provided that the diagnosis of dementia was made after the age of 60 years. We therefore excluded publications focused exclusively on early onset dementia as this is rare and typically has a different risk and disease profile compared to late-onset dementia [14]. Secondly, a risk model had to be reported along with performance indices (e.g., discriminative accuracy measured using the AUC or c-statistic [concordance statistic]). We used the following classification system to the interpret AUC/c-statistic: <0.70 = poor; ≤0.70 to <0.80 = fair; ≤0.80 to <0.90 = good; ≤0.90 to <1.00 = excellent [15, 16]. Thirdly, the outcome had to be dementia which included all-cause and its subtypes such as Alzheimer’s disease and vascular dementia (VaD) assessed using standardised criteria (e.g., Diagnostic and Statistical Manual of Mental Disorders: DSM). Studies where the outcome was a combined cognitive group, e.g., mild cognitive impairment cases combined with dementia cases, were excluded. Unlike our previous reviews, no restriction was made based on language.

The selection of articles followed the same protocol as our previous reviews [8, 9]. Firstly, four authors (J.B., A.H.K., R.K., B.C.M.S.) independently screened titles and abstracts for relevant articles based on the inclusion criteria. This was conducted using Rayyan software [17]. Secondly, the full text of any identified articles was checked independently by four authors (J.B., A.H.K., R.K., B.C.M.S.). Where duplicate studies were identified, details of all novel risk models and their performance indices (AUC/c-statistic and calibration statistics) were extracted from each publication and reported separately. Disagreements were resolved through discussion.

Two authors (J.B., L.G.) assessed the risk of bias of included articles using a modified version of the Newcastle-Ottawa Scale for non-randomized studies [18], the same adapted tool used in our previous review [9]. Excluding items that describe non-intervention cohorts, studies could achieve a maximum of six stars (compared to the original nine) based on selection, comparability, and outcome criteria. All scores were cross-checked by another author (A.H.K.).

An Excel data extraction sheet was created including the following information: author details, country, sample size, follow-up length, baseline age, outcome (e.g., all-cause and dementia subtypes; including number with dementia and diagnostic criteria used), the component (predictor) variables in each model, discriminative accuracy (AUC or c-statistic), calibration (e.g., Hosmer-Lemeshow test, Brier score, calibration slope, calibration intercept), and external validation metrics (e.g., external AUC, external calibration measures). Data were extracted independently by four authors (J.B., A.H.K., E.Y.H.T., B.C.M.S.). Any discrepancies were resolved via discussion.

As with our previous reviews [8, 9], a formal narrative synthesis was undertaken of all included studies. A table of all predictor variables included in each dementia risk prediction model was created to determine the key risk factors for incident dementia risk (all-cause and its subtypes). Using model performance indices (e.g., discriminative accuracy measured using the AUC and c-statistic), we created a plot for comparing the different AUC/c-statistic values of the different models, with the x-axis representing the AUC/c-statistic, and the y-axis representing the dementia risk model and study population. Different forms of c-statistic exist, and values are affected by factors such as population homogeneity and extent of censored observations, hence some caution is required when interpreting this plot. A meta-analysis was not possible due to little overlap in the nature of the predictor variables incorporated into the different risk models and large heterogeneity in populations studied, study design, and follow-up time.

In total, 9,209 papers were identified from the electronic search of which 4,376 were duplicates and therefore removed. After title/abstract screening of 4,833 records, 129 articles were identified for full-text review. From these, 51 were selected for inclusion. An additional 23 articles were identified from reviewing the reference list of included articles. Therefore, a total of 74 articles are included in the review (shown in Fig. 1).

Table 1 shows an overview of the included studies, with Online Supplementary Material 3 providing the complete data extraction. As shown, sample sizes ranged from 192 [19] to 930,395 [20], with baseline ages ranging from 18 [21] to 98 [22] years. Follow-up ranged from ∼1 year [23] to >30 years [24, 25]. Most studies used data from high-income countries (HICs; n = 67 studies), with a substantial proportion of these (n = 18 studies) using cohorts from the USA. Six studies only drew on data from low- and middle-income countries (LMICs) including China, Cuba, the Dominican Republic, Mexico, Peru, Puerto Rico, and Venezuela [26‒31]. Outcomes predominantly focused on all-cause dementia (n = 61) or clinically diagnosed AD (n = 19), followed by VaD (n = 1) [32], mixed dementia (n = 1) [33], Parkinson’s disease dementia (n = 1) [34], and Dementia with Lewy bodies (n = 1) [23].

Of the 74 included studies, most used a cohort design (n = 57), 12 used electronic health records, two used insurance claim data, and one used health survey data [55]. One study used a combination of administrative claims and electronic health records data [83], and another used cohort and electronic health records data [23]. Only 14 studies (19%) used population-representative data [27‒29, 43, 45, 66, 69, 70, 82, 85, 90, 92, 99, 100]. As noted in online supplementary material 3, most studies (n = 56) used traditional statistical modelling methods including Cox, logistic regression, and the Fine and Gray model. These different approaches estimate different outcomes, i.e., logistic regression estimates the odds of dementia as a binary outcome (yes/no), whilst Cox estimates hazard ratios, and Fine-Gray models (competing-risk) estimate sub-distribution or cause-specific hazard ratios. Models using supervised ML and AI methods have emerged, having been absent in both our previous reviews. This includes generalised linear models like Elastic Net [108], tree-based models utilising decision trees, random forests, and XGBoost [43, 44], non-linear models, for instance, long short-term memory networks [66, 97], as well as combined approaches ML/AI [19, 24, 28, 34, 37, 56, 60, 76, 77, 83, 86, 90, 96].

In total, 50 articles scored a maximum of six stars on the modified Newcastle-Ottawa Scale, whilst 19 articles scored five stars, four articles scored four stars, and one article scored three stars (online suppl. material 3). From this, it can be inferred that most studies included in this review have moderate or high methodological quality, with minimal bias observed in study reporting.

Of the 74 included studies, only 10 studies adhered to established reporting guidelines such as the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement [110] and the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines [111]. Specifically, nine studies followed the TRIPOD statement [22, 28‒30, 40, 51, 55, 72, 86] while one study adhered to STROBE guidelines [108]. This indicates that most studies did not utilise standardised reporting frameworks, which may impact the transparency and reproducibility of their findings.

Across the 74 articles, over 450 unique predictor variables have been incorporated within the context of dementia risk modelling, including demographic (n = 8), cognitive (n = 63), health (self-reported or objectively measured health status; n = 187), lifestyle (n = 13), imaging (e.g., magnetic resonance imaging; n = 35), genetic (n = 23), blood and protein biomarkers (n = 58), physical functioning (n = 25), and others (n = 55) including, e.g., multilingualism and healthcare utilization. Like our previous reviews, models range from incorporating single predictors [80, 98] to a maximum of 10,363 variables within a single model [83]. Figure 2 shows the 34 most used model variables; defined as being incorporated into a prediction model a minimum of four times across the 74 included studies. All factors previously identified in the Lancet Commission on Dementia Prevention, Intervention, and Care [112] were frequently used as model variables, except for air pollution. Most variables are modifiable except for age, genotypic sex, ethnicity/race, and genes (e.g., apolipoprotein E e4).

Overall, 63 studies developed new dementia risk models since our previous reviews [8, 9], with 21 studies developing models without undergoing any validation procedures, 15 reporting both development and internal validation metrics, and 27 reporting internal validation results only (Table 1). This is a notable increase in the number of studies that have undertaken internal validation, with cross-validation resampling most frequently used.

Discriminative accuracy of models for all-cause dementia ranged from an AUC/c-statistic of 0.50 (logistic regression model using latent scores of hearing difficulties; follow-up = 10 years) [45] to 0.96 (discrete Bayesian networks developed model incorporating age, mini-mental state examination score, and fish olfactory identification; mean follow-up = 4.9 years) [48]. For AD, discriminatory accuracy ranged from 0.54 (ML incorporating numerous laboratory tests, healthcare data, and sociodemographic features; follow-up = 10 years) [96] to 0.92 (logistic regression using a Free and Cued Selective Reminding Test Risk Score; follow-up = 3 years) [113]. Developing a model exclusively for Dementia with Lewy bodies, one study found that the number of core clinical features, neuropsychiatric symptoms, and cognitive score affects discriminative accuracy (range from 0.56 to 0.90; follow-up = 1 year) [23]. One study incorporating polygenic risk scores into basic models reported fair discriminative accuracy in predicting VaD (AUC/c-statistic = 0.78; follow-up = 17 years) and mixed dementia (AUC/c-statistic = 0.84–0.85; follow-up = 17 years) [32], whilst another study utilising cognitive-based variables reported fair discriminative performance in predicting Parkinson’s disease dementia (AUC/c-statistic = 0.77+; follow-up = mean of 4 years) [34]. As expected, internal validation was often fair (AUC/c-statistic between 0.70 and 0.79) or excellent (AUC/c-statistic ≥0.8), although four ML-based models reported poor internal validation results (AUC/c-statistic <0.7) [60, 83, 96, 97].

In total, 27 dementia risk prediction models were externally validated (Fig. 3). This included models specifically aimed at developing a score for dementia risk, e.g., the Brief Dementia Screening Indicator (BDSI) [49], the Cardiovascular Risk Factors, Ageing and Dementia risk score (CAIDE) [52], the Australian National University AD Risk Index (ANU-ADRI) [73], Study on Aging, Cognition and Dementia (AgeCoDe) [94], and Lifestyle for Brain Health (LIBRA) [91, 107], or models that were developed to index risk in cardiovascular disease, such as the CHADS₂ [71], CHA₂DS₂-VASc [58], and AHEAD [63]. As shown in Figure 3, model discriminative accuracy and calibration varied greatly across the external validation cohorts, but overall was generally poor. Models that transported well include the Basic Dementia Risk Model (BDRM) [74], BDSI [49], and Chouraki et al.’s [42] model. Other models such as Ben-Hassen et al.’s [38] and the Rapid Risk Assessment of Dementia (RADaR) [40] reported high discriminative performances but need further external validation to assess their transportability. Discriminative performance across models not specifically aimed at dementia risk was generally poor.

Only 20 studies reported calibration results. Studies that developed new models generally reported good or excellent calibration [50, 61, 65, 69, 85, 96, 99], with few studies reporting mixed results with calibration ranging from poor to good [55, 90], or poor [37, 78]. Where previously developed models were tested outside the context in which they were developed (i.e., externally validated), the results varied from having poor [51, 72] to mixed/good calibration [20, 30, 22].

Of the 74 included studies, 16 developed risk models in Asian populations using data from Japan (n = 7), China (n = 3, South Korea (n = 3), Taiwan (n = 1), and Hong Kong (n = 1). Only five studies developed models in LMICs, using data from Mexico [26, 27] and China [28, 48, 31].

Ethnic variability in predictive accuracy of risk scores has been investigated within African American, Caribbean, Asian, and Hispanic populations residing in the USA [40, 42, 47, 50, 84]. Developed in a predominantly White cohort, the RADaR model was externally validated in the Minority Aging Research Study (MARS; 100% Black cohort) [40]. With a prediction horizon of 3 years, the RADaR displayed excellent discriminative accuracy (AUC = 0.85) that was higher than the development cohort (AUC = 0.81), with baseline characteristics varying greatly, i.e., MARS participants having, on average, higher prevalence of diabetes and increased body mass index. Four studies explored dementia risk in Mexican [26, 27, 30] and American-Mexican [50] (participants identifying as Mexican and residing in the USA) cohorts. Using data from the Mexico 10/66 cohort, one study developed a model with “fair” discriminative performance (AUC range: 0.72–0.75), utilising variables such as neuropsychiatric symptoms, rurality, illiteracy, mild cognitive impairment, and diabetes, with participants followed for 3 years and aged >80 years [26]. To note, this does not mean it is suitable for applications to unselected populations in LMICs. As shown in Figure 3, another study externally validated various dementia risk models using data from the 10/66 study over a 10-year follow-up [30], reporting that the BDRM, BDSI, and ANU-ADRI transported well into Mexican populations (c-statistic = 0.71), whilst transportability of the AgeCoDe and CAIDE was poor (c-statistic = 0.64 and 0.58, respectively). Another study developed the Mexican American Dementia Nomogram (MADeN) risk index that aims to predict dementia in Mexican Americans over a 10 years period [50]. The MADeN risk index included education, diabetes, age, and other novel measures such as pain and social support/engagement. Overall, the MADeN exhibited fair predictive accuracy (AUC = 0.74) with good calibration.

Six studies have developed and tested models in disease-specific groups including: type 2 diabetes (n = 2) [69, 78], cardiovascular disease (n = 2) [57, 62, 70], and stroke (n = 1) [21] populations.

One study developed a model incorporating; age, sex, diabetes duration, body mass index, variation (%) fasting plasma glucose, HBa1c, stroke, hypoglycaemia, postural hypertension, coronary artery disease, and anti-diabetes medications in a sample of n = 27,540 type 2 diabetic patients in Taiwan, aged 50–94 years, with a mean follow-up of 8 years [69]. The model had fair/good discriminative accuracy in the derivation (AUC range: 0.76–0.82) and internal validation (AUC range: 0.75–0.84) datasets; with shorter follow-up times having higher predictive accuracy scores (i.e., 3-year vs. 5-year and 10-year follow-up). Another study developed and validated the RxDx dementia risk index, incorporating demographics, the Charlson comorbidity index, and a chronic disease score specifically for older (≥60 years) individuals with type 2 diabetes and comorbid hypertension [78]. The RxDx risk model demonstrated excellent discriminative accuracy in both the training (c-statistic = 0.81) and validation (c-statistic = 0.86; type 2 diabetic patients without hypertension) datasets. However, calibration was poor.

One study investigated risk of incident dementia using the CHA₂DS₂-VASc and AHEAD scores in a large national database in Taiwan of n = 387,595 participants with heart failure [62]. Although high CHA₂DS₂-VASc and AHEAD scores were strongly associated with an increased risk of dementia over a mean of 2.9-year follow-up, model discriminative accuracy was poor (AUC range: 0.55–0.61). In the same dataset, but focused on participants with atrial fibrillation (n = 332,665), CHADS₂ and CHA₂DS₂-VASc scores again showed poor discriminative accuracy (c-statistic range: 0.59–0.61) for predicting incident dementia [70]. Another study looking at dementia risk in patients with atrial fibrillation (n = 74,081; followed for 10 years) in the USA, reported a higher predictive accuracy of the CHA₂DS₂-VASc (AUC range: 0.65–0.73), compared to the Intermountain Mortality Risk Score (IMRS) (AUC range: 0.68–0.69) [57].

A dementia risk model incorporating demographics, the Charlson comorbidity index, and cognitive symptom burden showed poor discriminative accuracy (c-statistic = 0.59) for incident dementia in stroke patients (n = 18,681) over 5-year follow-up [21].

Research stratifying sex frequently showed analogous discriminative performance. Over a 10-year observation period using common demographic, health, and lifestyle variables, Park et al. [85] reported consistent discriminative performance across male, female, and aggregated participants in both development and internal validation cohorts (n = 331,126 using Cox regression; AUC = 0.81) with excellent calibration, as reported by both calibration slope and intercept. Two studies used Fine-Gray sub-distribution hazard modelling methods, reporting fair or good discriminative ability across sexes over 5 years (n = 75,460; c-statistic = 0.83 for both men and women) [55] and 10 years (n = 2,710; c-statistic ranging from 0.79 to 0.87 for men, 0.78–0.85 for women) [74]. One study reported poor discriminative accuracy across both sexes in their 10 years, 42 health and functioning variable model (n = 7,239; AUC = 0.67 for men 0.66 for women) [92]. One study using the CHA₂DS₂-VASc and IMRS models to predict 10-year dementia risk in atrial fibrillation patients (n = 74,081) found discriminative performance varied more in men (AUC range: 0.27–0.68) compared to women (AUC range: 0.59–0.77) [57].

In this review, 74 studies were identified, adding to the 46 studies already published. Since our previous reviews in 2010 [8] and 2015 [9], there has been an increase in the development of novel dementia risk prediction models. While traditional analytical methods are still being used, more advanced computational methods (i.e., AI and ML), population-specific models targeting groups with higher risk of dementia based on disease comorbidity, such as individuals with type 2 diabetes, atrial fibrillation and stroke, and a greater focus on validation (both internal and external) is now taking place. Further, for the first time, models have been developed and tested in LMIC settings. However, model performance indices, captured in the AUC/c-statistic, remain variable (0.50–0.96) and not all models show good transportability/generalisability, particularly when external validation is tested. Moreover, calibration was rarely undertaken, making it difficult to ensure the predicted probabilities of developing dementia align closely with the observed probabilities, and whether risk models may be under- or overestimating risk. Although some models show acceptable transportability, it is still not possible to recommend one model for use in clinical or population-based settings without further robust validation, as highlighted by other reviews [9, 114]. However, there are several models that show promise, such as the BDSI, BDRM, and RADaR that could be the focus of further investment, as the proliferation of developed risk models may further fragment the evidence base.

The inconsistent adherence to established reporting guidelines such as TRIPOD and STROBE raises concerns about the completeness and transparency of the reported methods and results. Adherence to guidelines like TRIPOD and STROBE is crucial for ensuring that studies are reported with sufficient detail to allow for critical appraisal and replication. Future research in this area should aim to adhere to these guidelines to enhance the quality and reliability of prognostic models. Below, Table 2 outlines standards risk prediction should be held at, in addition to pre-existing guidelines.

Demographic, health, and lifestyle variables are often incorporated into dementia risk prediction models. This resulted in substantial overlap between model components, e.g., the 12 factors outlined in the Lancet Commission on Dementia Prevention and Care including: early life low educational attainment, midlife obesity, alcohol consumption, and hypertension, and later life diabetes and smoking [112]. Importantly, the risk modelling literature expands on these 12 factors and includes over 450 predictor variables such as cerebrovascular diseases (e.g., stroke), cardiometabolic conditions and their risk factors (e.g., hyperlipidaemia, heart failure, atrial fibrillation, and coronary heart disease), genetics, and cognitive function (e.g., verbal fluency).

The use of ML/AI methods in large datasets has allowed for more intensive testing of a larger number of variables. Indeed, in one study over 10,000 predictor variables were used in the training model matrix with 50 predictors incorporated into the final model [83]. This exploration of the number of predictors tested could potentially help identify new mechanisms of disease and inform the development of novel risk reduction and prevention strategies. However, it is crucial to evaluate whether ML/AI models outperform those models developed using traditional methods and more restricted risk predictors. Specifically, the advantages of a greater number of predictor variables used in ML/AI models are uncertain when balanced against the time and cost of obtaining data, overfitting and the consequent lack of transportability, and model predictive accuracy. Indeed, our review highlights that most AI/ML models did not outperform those developed using traditional statistical methods.

As in our previous reviews [8, 9], some models include time and cost-intensive variables such as blood-based biomarkers, genetics, cerebrospinal fluid analysis, and brain imaging. However, as highlighted in one study such information may only add limited predictive value at huge infrastructure and personnel costs [93]. It is also possible that different risk factors and the way they are combined (or scored) should consider the setting in which risk modelling is being undertaken. Indeed, an attempt to transport models developed in high-income country settings to LMICs in one study showed poor external validation for most models tested with the exception of the ANU-ADRI, BDSI, and BDRM [30]. Different risk factors may require testing for LMICs (e.g., illiteracy, food insecurity, and infectious disease profiles) which may be less applicable to high-income settings.

In studies that have considered the impact of ethnicity; namely African American, Caribbean, Asian, and Hispanic populations in the USA [40, 42, 47, 50, 84], it has been observed that olfactory variables, may require adaptation to be culturally appropriate for the studied population [47]. In contrast, the RADaR model, incorporating common demographic, health, cognitive, and genetic, and functioning variables, showed similar performance in White versus Black populations [40]. Risk models either developed or validated in Mexican populations (also residing in the USA) often showed fair predictive discriminative performance, such as the MADeN model that was exclusively developed for use in Mexican populations [27, 50], as well as the ANU-ADRI, BDSI, and BDRM [30]. Although very few studies explored sex differences in dementia risk, the majority that undertook this demonstrated comparable results, with both sexes presenting a similar level of discriminative performance suggesting robust reliability and validity across both sexes [55, 74, 85, 92].

There has been an increase in the number of models developed in disease-specific populations including individuals with stroke, atrial fibrillation, heart failure, and type 2 diabetes. These typically incorporate risk factors identified in the whole population as well as disease-specific variables such as anti-diabetic medication, hypoglycaemia, and renal parameters. Parallel to whole population findings, the accuracy of disease-specific models varies. Generally higher accuracy is observed in cardiometabolic-specific versus stroke-specific models. Difficulties with predicting post-stroke dementia could be due heterogeneity in stroke subtype, impacted brain region, pre-stroke cognition or other yet undetermined risk factors specific to this population, all of which could differentially influence risk models.

One study tested whether models developed in the general population transported well to stroke populations. The results showed poor performance across all tested risk models, indicating that future work is needed to develop dementia risk prediction models specifically in stroke populations [115]. Two studies investigated whether scores originally developed to predict long-term mortality in European heart failure patients or stroke risk in atrial fibrillation [62, 70] were predictive of incident dementia in heart failure and atrial fibrillation populations. The models also showed poor discriminative performance. Overall, the results highlight the need to develop dementia risk models tailored to specific disease populations [116, 117].

The review has several strengths. Firstly, the systematic approach, using the same protocol as our previous reviews, ensures that all relevant literature was captured, and the results are comparable across publications. Secondly, no language restrictions were applied thereby reducing the bias associated with using articles only published in the English language. There are also some limitations. Due to considerable heterogeneity across studies, e.g., in the predictor variables selected, follow-up times, dementia outcome use, data used, analytical methods, and sample demographics, only a narrative synthesis of the results was possible. Further, few studies used population-representative data thereby limiting generalisability of the results. Limited testing across settings (high-income country vs. LMIC) and subgroups within society (e.g., stratified by ethnicity, sex and socio-economic status) also raises questions around transportability/applicability. Unfortunately, we were unable to report some model performance metrics where authors presented the data graphically, preventing determination of specific figures. It is also important to note that while some studies reported good calibration, calibration assessments based solely on the Hosmer-Lemeshow test may be misleading. As such, calibration should also be evaluated using calibration plots or tables comparing predicted versus observed outcome frequencies, as the Hosmer-Lemeshow test alone has limited suitability for assessing calibration, particularly in large datasets [118].

While the growing number of dementia risk models reflect the increasing interest and investment into the area, our systematic review shows that there is still no one model suitable for use in routine clinical practice or the general population. It is essential to now validate a limited number of high-quality, context-specific models to determine whether they are useful if implemented in diverse populations. Once the clinical validity and effectiveness of risk prediction models are confirmed, further analyses on cost-effectiveness and clinical utility should be undertaken. This includes evaluating potential harms, patient, and public acceptability of such tools, the implications of any such effort for health systems to address questions around who would be responsible for undertaking scoring and risk communication, and how model outcomes could inform “personalised medicine” plans [119]. Moreover, population-based strategies, in tandem with personalised medicine approaches, may provide a more holistic strategy to preventing dementia by not only having the potential to significantly reduce the overall prevalence of dementia but may also help address health disparities by ensuring that dementia prevention efforts are built in to society, and those that are individually orientated are appropriate and accessible [120, 121]. Finally, there needs to be a consensus on the most well-developed models and focussing research resources on improving the clinical validity and utility of these models as a clinical risk prediction tool for application across diverse communities in the general population.

A statement of ethics is not applicable as this study is based exclusively on published literature.

The authors declare no conflict of interest.

This work was supported by the NIHR (Three Schools: Dementia Research Program, Improving the lives of people living with dementia and carers 2021–2024; l Grant reference #DP06) and UKRI (Research Grant: MR/X005437/1). The authors declare no conflict of interest.

R.K.A., M.B., C.B., G.F., L.R., M.S., E.P., C.M., D.R., N.Q., and B.C.M.S. conceptualised and helped in the funding acquisition. L.E. refined the search strategy. J.B., A.H.K., R.K., and B.C.M.S. undertook title, abstract, and full-text screening. J.B., E.T., and B.C.M.S. developed the methodology. J.B., E.Y.H.T., and R.K. performed data extraction. J.B. and L.G. undertook the risk of bias assessment. J.B. curated the data, undertook project administration and data synthesis, and prepared the original draft. B.C.M.S, J.L., A.S., P.J.T, and D.T. supervised the project. All authors reviewed and edited the final manuscript.

All data supporting the findings of this study are available within the article and its supplementary materials. Further enquiries can be directed to the corresponding author.